Latest / The Bleeding Edge Podcast / Weekly AI Brief July 03 2026
Transcript
- Ralph: Hi, and welcome to another episode of the Bleeding Edge Podcast. I'm your host, Benka, and my co-host in the studio with me today, Emil Emil, welcome to the Edge. Emile: Thanks, Ralph. Good to be back. Ralph: Yeah, we've had a little hiatus, ⁓ travel plans and work, and a few things have meant that we have had to be elsewhere. But we are back this week, and of course we have two weeks worth of news to catch up on, which is obviously gonna take us a bit of time. But everything is on the hyphen edge, hyphenai dot versal dot app, and ⁓ you can catch ⁓ on of the stories that we have been gathering in the news ⁓ and about in the world of AI. And of course, ⁓ there are some very exciting updates. probably start with the first one that Mythos is in the form of Fable Five, but some say with its balls cut off, a slightly somewhat muted Fable Five has made a reappearance. Emil, I'm sure you will share some of the deficiencies ⁓ people talking about on X and some of the other platforms. but it's back and actually has been doing some really good work for me. It's going to be kicking around, ⁓ available to us ⁓ on ⁓ plans until the 7th of July. Unfortunately, no free usage, but we you are able to use up to 50% of your available spend of your available tokens you can now actually you know use with Fable 5, which is good news. And on to a couple of the other stories. So it seems that it's all about chips and memory and hardware. ⁓ in the last week. So what we've seen is the memory squeeze is becoming AI's binding constraint. So really interesting article in the FT. that ⁓ Apple ⁓ get this, love this story, is actually lobbying the US government to buy memory chips from blacklisted Chinese suppliers to cover the shortfall. I think that's incredible. Emile: Wow. Ralph: so you know there you go that what to do with your trade a deficit, go forth and buy chips from blacklisted Chinese suppliers. and so we've also seen that ⁓ micron signed a investment deal with Anthropic for chips and OpenAI detailed its first custom chip so We are moving from software into again the hardware era. It's funny how these things seem to go around in cycles. but ⁓ you know, what have we had? We've kind of had the software, then we had the data centers, and now we've got the chips. Literally everything ⁓ ⁓ you ⁓ converging and is becoming a pinch point. So capacity is literally straining the entire supply chain. But all of the labs ⁓ in of the the news what it looks like is they are racing to lock up and DRAM. ⁓ those are the kind of key hardware components. I can talk a little bit more about the roles of those but ⁓ you know that has become the significant barrier. So there are fifty six US firms that have been banned by China. So, you know, this is kind of going both ways, mostly defence contractors, rare earth miners. and that is relating ⁓ for the the Pentagon military ⁓ blacklist. and the reason all of this is getting very exciting is because the constraint is shifting from mod models to the physical substrate. of course, no doubt, you know, back in the news will be we've run out of water and we've run out of energy and electricity. but at the moment it's very much on the hardware side. So ⁓ obviously the US and China fighting very much a a war of attrition on who can ban the most firms the fastest. ⁓ moving on to another interesting news item the last week from OpenAI. They launched daybreak. So what they've now got is their kind of a methos equivalent GPT 5.5 Cyber and they have launched an initiative called Patch the Planet. So no doubt this is in response to obviously Fable 5 ⁓ being pulled because ⁓ essentially allows anybody with a keyboard ⁓ the ability to do sophisticated y you know attacks let's just say and you can expose vulnerabilities ⁓ openAI have basically chosen to you know from the front and say this isn't about exposing vulnerabilities and hacks, this is actually about patching vulnerabilities first. obviously ⁓ what is going now is we're kind of seeing that the most powerful models are being released to a small number of actors initially. ⁓ you know, the idea there is that they experiment with those models, identify the vulnerabilities, patch them, hopefully share them with ⁓ all the good actors in the world before the bad actors actually get their hands on this capability. And so so this is kind of, you know, fascinating. I think it it's kind of it's it's kind of Changing the idea of defense, right? So ⁓ we're we're really seeing that ⁓ our most vulnerable ⁓ platforms are in fact software platforms. And so actually, you know, war is maybe something of if we think of it in this way. Obviously that's not what's going on in the Ukraine and elsewhere, but ⁓ certainly for the rest of us this is a much bigger reality than somebody you know, raining down bombs. And so I think fixing the flaws is definitely an interesting way of tackling it. But I am a little bit concerned that it's no longer, you know, the democratization. Can I say that right? The democratization of you know the AI capability, knowledge for all men. no, it's knowledge ⁓ an application of AI for the few. And maybe you know we'll get what's left over. So ⁓ the model wave also rolls on in the news. We are hearing about GPT five point six Soul, which I'm not quite sure if it has been fully released, but I know it's in preview. And ⁓ five point six is basically pushing ⁓ reasoning even So other releases that seen, Deep Seek V4 plus D Spark. Fable Five, which appeared sometime on a late evening in the last week ⁓ on my available list and I started to use it immediately. and so it's back from the dead. and hopefully you you can tell us how it performs because I know you shared a couple of very interesting visualizations on that. so ⁓ just jumping into a little bit about Deep Seek, obviously Deep Seek is a Chinese LLM. a ship D Spark and it is a speculative decoding method and it basically runs Deep Sea fifty seven to eighty five percent faster than its prior version. And so that ⁓ is exciting. Deep speed goes pretty fast even before all of that. but you know kind of seeing full stack model refreshes from three labs in one week like it's it's pretty impressive. we know things were moving fast, but this kind of release cadence, you know, things are not cooling down. We are just seeing tremendous strides, even in a two years into what was already a accelerated journey ⁓ amongst many different labs. And so I think ⁓ we've got our work cut out for us, Emil. one other ⁓ item here that caught my eye this week ⁓ an interesting story about a guy called John Jumper who has jumped. he's leaving Google DeepMind for Anthropic. So we've seen Andre Karpathy, who was obviously OpenAI. He was also in Tesla. ⁓ he founded his own AI education company. ⁓ then at some you know went back into open AI. and now he's joined ⁓ And now we've got Jumper. who is a Nobel laureate for basically creating the ⁓ Alpha Fold routine which if memory serves me correctly is able to synthesize proteins so this guy ⁓ is a real pant and ⁓ he he his stuff and so anthropic are building the A team not only are they the only lab that is profitable And not only are they the most valuable lab, but still only eight percent of the market share, they are now also ⁓ you know building ⁓ the team. And it seems to me that people are joining them for the mission, which is kind of interesting because that's the only way that you can really win the talent war. but essentially DeepMind have just lost their most decorated researcher to a direct rival. So that ⁓ is Going to hurt, and we'll have to see, you know, what is going to cook up at Anthropic. and then a couple of other items just back on the hardware side. ⁓ the labs are definitely buying their way into silicone, and obviously, know, we've kind of ⁓ already spoken about the micron anthropic deal, OpenAI's first chip. but ⁓ Dell has also released a GB10. Desktop, and so have not quite checked out the specs on that one, but it is reported as being something that you can essentially run your local model on, and as we know, local is lacquer. having a local model ⁓ is something that should all aspire to if we don't have to spend too much money. So Dell shipping a workstation built on Nvidia's GB10. a big deal because it's kind of signals, you know, accessible hardware that doesn't cost you ⁓ tens thousands, but something that you can actually afford and build your own home lab. So ⁓ those were the top stories ⁓ of last week. Anything interesting there for you, Emil? Emile: Yeah, obviously, Fable 5 is the big story of the week. So just want to add a little bit more color around that. As you've mentioned, Rolf, they redeployed globally after US export controls were lifted, whereby they essentially have added stronger classifiers to block cybersecurity misuse while promising refinements to minimize the false positives. Initial user engagement on X is suggesting that those sensitivities are set quite high at the moment. they'll iron that out. But ⁓ a bit of a mixed bag, Rolf, like what I'm experiencing, one end users are complaining saying it's been heavily nerfed and others that, you know, have more sophisticated harnesses set up and workflows ⁓ sailing through and reporting good things. So I guess it's just really how you ⁓ the model, you know, those controls are there. Maybe there's ways around it or maybe it's just the specific use cases each person is using and each project is giving a different set of results. But yeah, it's, mean, the overall sentiment is people are frustrated. They've lost time on it, but equally, you know, everybody's just happy that it's there and they can use it again. yeah, it'll be interesting. The routine coding and debugging tasks will temporarily fall back to Opus 4.8 under the new safeguards. The announcement includes plans for an industry framework also with Amazon, Microsoft, Google, and others jailbreak assessments plus deeper US government collaboration on model testing and safeguards. Audience reaction to Anthropix announcements is generally negative and frustrated, as I mentioned. Users widely view the new classifiers, fallbacks to coding and debugging to Opus 4.8, short availability window and usage restrictions as significantly diminishing the model's value for paying subscribers. A small minority acknowledge legitimate safety and misuse risks, especially around cybersecurity, but they criticize execution is overly restrictive for legitimate use cases and write and 30 so many express intend to reduce reliance on entropic or switch providers citing that some routine tasks like coding and debugging where entropic's actual words is broadly seen as misleading giving that coding is a core high-value use case for frontier models There is a broader theme of eroded trust. And I think this is really the main story here with calls for migration to open AR models or open source alternatives. example of common themes that I saw in replies captured it quite well in terms of frequency of comments I saw was, so we basically have fable on our sub for less time than we originally planned for less usage allowed than the sub originally allowed. And it also can't be used for coding tasks during the time we can use. it question mark. But either way it's great to see the model out there so we can use it. we get lucky and they do another extension but yeah enjoy it while you have it. Ralph: Yeah, and then of course they actually came out with ⁓ Sonnet ⁓ five as well. So kind of double stacked ⁓ the model release. ⁓ I have been using Fable Five quite successfully and a top tip, I've been obviously sharing this in our various groups, but one of the best ways of using Fable Five is actually building a that is able to architect prompts. For Fable 5 to use. And the reason why that is so effective is because you can really build in the limitations of 5. So basically the prompt architect that I put together in a skill checks what Fable 5 likes to be prompted. And also ⁓ not to those gotchas, so those guardrails where it then Either stops processing or it degrades to a lower order model. And that has actually worked very successfully for me. So I have been seeing it do very long coding runs. But ⁓ only you know when it's actually been a task that didn't require fable five. So ⁓ I I highly recommend ⁓ going in and just asking to build a skill for how to prompt you know each of these models and you can just build one skill and train it on the best prompting techniques for each model. And of course you can go to the official GitHub Emile: you Ralph: which ⁓ tropic has and on there it sets out exactly how each model wants to be prompted, drop that into a skill. And every time you start using Fable Five, you start with prompt architect, you write the intent and get it to write the prompt and you have a much better experience with Fable Five. And you also really get it to ⁓ optimize the model does. So Fable Five is particularly good at long runs. ⁓ also what we saw when it initially came out, it does stuff that it's not been asked to do but knows is fundamental to your mission. So the other top tip is ⁓ with the outcome. Start with what you want the user experience to be of the piece of software or solution is that you're building, and you get the prompt architect to actually build a prompt to deliver that outcome. And with that, you get a much broader instruction for Fable 5. ⁓ that means it will go through a design process, a build process, even a testing process. Obviously, debugging ⁓ one of those gray areas where you might still hit some issues, but I've had it, you know, test and fix the applications that it's actually built, ⁓ and it's been very successful in that way. So I think it's down to you know partly how you actually what you're asking the model, how you're asking the model. And there is no way we have the time to learn how to prompt each of these models. So just go forth and prompt your intent. forget about ⁓ you know context engineering and prompt engineering. You're now in the world ⁓ intent engineering. Explain the outcome that you're looking for and will build the rest around that. that's my top tip for for how to use it. Emile: Thanks, Rolf. That's really good advice. And just to compliment the tips you mentioned, I saw an interesting post from Leo. He's also been getting a ton done with Fable yesterday and he's not hitting his rate limits. he wanted to share some tips in, know, how is it achieving that success rate. He's using Fable on high effort for now. ⁓ X-Hi is is really token hungry and MaxX Ralph: But ⁓ Emile: also Claude code, how to use codecs as a fallback for lots of implementation tasks. ChatGBT is steerable and Fable can learn how to steer it. His third point is his road up, road type of big section in his Claude MD file for how to prioritize different models for different work when orchestrating workflows and sub-models. agents his fourth point is things that are ⁓ token hungry like computer use and code base analysis. He'll do with other models and report the results back to Fable. He does admit that he hasn't read much of his own MD file just via coded it out and what was looking for with Fable ⁓ and it confirmed. It can use Codex for things he cares about. So, you know, as long as it does that, Bob's your uncle and still finds Codex to be way better at computer use, verification for UI and UX work and generally more efficient than at execution on well-specced work. So in minimaxing, this has generally been really fun for him and loving the outputs that he's been getting. was, he's saying he was throwing away 50 % of his end-to-end agent driven PIs before building this workflow and he hasn't had to close a single one since yesterday. So pretty positive. Ralph: Yeah, sounds like he has out a very effective pattern and I love the idea of ⁓ guiding the model ⁓ well, just guiding the LLM on which model to select. ⁓ So an engineering standpoint, ⁓ I do models, but not as often as ⁓ I would mainly just because ⁓ it it takes time to do that extra step and and actually steering model. ⁓ is is a really smart way of doing it. So I'm definitely going to borrow that tip and I will check that out. but you know your point there, ⁓ Emil about not becoming dependent on any one model. I think that has been the big shift for the market over the the last couple of weeks. people know that, you know, just because the model is there doesn't mean it's actually going to be made available to them. we are not in a free market economy anymore when it comes to these very capable AI models. It is now very much ⁓ political endeavor. ⁓ obviously, you know, at the end of the day, ⁓ the models we do have access to are so capable. but you know, I think we definitely need to create options. and that is something that I'm working on also in my own work. but it's really difficult. So ⁓ it takes a lot of time to be able to set up access to all these different models. you know, even though it's a well trodden path, it's ⁓ not necessarily something that we always have a lot of time for. As ever, the human in the loop ⁓ is biggest point of friction. So ⁓ and of course we are the humans, ⁓ as far as I know. But that's really interesting news, and maybe actually it might be a good Emile: to you. Ralph: Opportunity just to share a little build that I actually did with ⁓ Fable Five. ⁓ don't have a lot of time, unfortunately, for our pet project ⁓ which we have been reporting on on the Bleeding Edge Podcast over the last few months. ⁓ but of course, ⁓ one of the ways that I am proving what AI is capable of ⁓ by building out ⁓ a first record label. and is now you know going through various iter iterations. Why have I picked this project? Well, I think it demonstrates obviously the music generation capability of AI. ⁓ it also demonstrates how to actually put workflows together, complex workflows in order to ⁓ orchestrate content, and then also actually how AI can interact. With various systems like YouTube and SoundCloud Spotify, all of these applications that typically host our music. And it also speaks to ⁓ the whole licensing debacle about AI-generated content, because ⁓ as we know, Warner Music, Sony Sony Music, Universal Music, they have all been in extended legal battles with the likes of Yu Dios and Suno. our know ⁓ legal is constantly shifting. in addition to that, it also demonstrates the capability ⁓ of e-commerce this particular music label has ⁓ access to a merchandise channel for each of the different brands. And it also speaks to the generative capability because ⁓ as part of this music generation, I have also put in the pipeline that it generates still images and videos. So I felt like this was the most complete way that we could really test a whole slew of technologies in an extended project. And of course, the aim is to see can you actually get ⁓ an AI native end-to-end business up and running, have it run ⁓ autonomously with a human in the loop where it matters. And and so without further ado, you know, would you like to see ⁓ what it looks like? ⁓ Emile: Please, pick me up, come on wait. Ralph: Great. So we go. ⁓ hopefully you can see this okay. ⁓ I will just zoom in. So ⁓ me just give some context here. ⁓ this ⁓ a website, it's actually a web application that was developed by Fable 5. ⁓ it ran for about 20 minutes and produced this. And one of the things it did was it created a radio station. So this particular music label, San Carlos Records. now has its own radio station. ⁓ you can see we've got the Alpacas Inibitha theme song, which is an album I wrote for a local alpaca farm. ⁓ and ⁓ the theme song is really introducing the soap opera that we have also ⁓ written, but not yet filmed. ⁓ but there it is. These alpacas are gonna be truly famous, ⁓ and ⁓ radio station is really effective. I I I really love it. Fable five did a great job. Integrating all of that with SoundCloud. You can actually see here are each of the personas. There's a really cool little action that it's got as you move through each of the albums. It looks very professional. All of this ⁓ built with Fable 5. I did have an earlier mock-up that was actually built with Sonnet 4.8. Emile: and arts. Ralph: and so it wasn't that Fable Five started with a blank canvas. There was an awful lot already set out as a mock-up that I did about two months ago. ⁓ when I found 20 minutes ⁓ in the last week, I just said, right, Fable Five, go and have a look at this and optimize it. You can see we've got a couple of issues here, just ⁓ with a few things that are not lined up correctly, but that'll be a quick fix. ⁓ and it's then gone forth and really set out. the ⁓ business model for this autonomous ⁓ music label and you can go in and license ⁓ a particular record. ⁓ you can actually ⁓ well basically you you you can license it to performance to perform a song ⁓ you can license it to produce your own song from you know a a sound you can actually license a particular song you can license a character or you can commission your own. So those are the models over there, and it's come up with some prices. ⁓ so ⁓ you know, I am essentially giving the ⁓ free reign to figure out what an AI native music label should contain, how it should monetize, ⁓ and through my prompting, encouragement, and nudging, ⁓ it's basically set out this ⁓ model. and then obviously you can see over here that you have the option to basically build your own AI first label and get everything in the box. And everything in the box is ⁓ a tool that ⁓ so there are 23 agents sitting behind this thing. ⁓ everything the CEO to the ⁓ lyric writer to the ⁓ actual ⁓ marketing agent. ⁓ it means that ⁓ you know it starts with basically building a persona, so you know, who is the artist behind ⁓ album, what is the reason for the album, what's the inspiration, ⁓ then basically it kicks off a catalogue pipeline, which is then able to ⁓ write the songs ⁓ build the albums and ⁓ the way through to actually creating the artwork. ⁓ and then where the human in the loop comes, obviously listening to what it produces, curating the list, throwing out anything that is not at the right level of quality. And then I have to manually ⁓ publish to SoundCloud. So one of the things I'm finding, and this is a control point by the music industry, ⁓ is that essentially a manual upload is required. ⁓ there are certain things that you can actually get the AI to do through. Computer use, but actually, in terms of the terms of service, you have to manually upload these things into the likes of SoundCloud and Spotify. So that's all fine. It means that there is real human curation at play. And ⁓ a bit of a shout out to Jack Matthews, who said when England will win the World Cup, you have to release a song. And so one of the Tests that I have done is I have actually had the pipeline generate the album ⁓ for England and it goes through and there's a song for each match, and obviously it basically sings about the context and the experience of that match. ⁓ and it has matched the genre to the team that they were playing. So obviously ⁓ each song has a slightly different flavor. We'll have to I haven't actually generated the music just yet, but Emile: That's cool. Ralph: The album is written, the lyrics are written, the ⁓ sound profiles have been picked. ⁓ and this thing goes through and masters all of the songs when they are produced. ⁓ and then eventually what it will also do is it will apply for any copyright that it can. So the lyrics can be copyrighted. and ⁓ essentially what you then have is a real asset in terms of. the music that you've generated and you can then turn that into merch. You can also have the album written to vinyl. ⁓ think r the cost of that is around ⁓ 30 euros ⁓ twenty twenty eight pounds thirty nine dollars ⁓ you like and so ⁓ it's very accessible to actually have your music put on vinyl ⁓ ⁓ essentially you know y you can buy this entire platform which is the label in the box ⁓ the AI has set A very accessible price of 18k for that. There is a lot of tech behind it. ⁓ And so if you have 18k to spare and you want to generate your own music label and have it generate the music through a pipeline and basically run through the full distribution and marketing process in addition to driving merch sales. integrated with a Spotify e-commerce catalog that is coming soon as you can see over here. ⁓ it will then also have the full integration into the various ⁓ media platforms for marketing. So ⁓ think you know what we're heading to here is going to be a digitally ⁓ well a distribute distributed automation ⁓ organization what they call a DAO this thing should run reasonably hands-off, other than those critical points where human creativity has to really come in to review the lyrics and approve them, and then obviously do the final selection of what music actually goes forth. And you can kick this whole thing off ⁓ by essentially sending a briefing ⁓ prompt through on Telegram. ⁓ And Telegram just connects to the AI agent. the AI agent will ask you a few questions and then it will basically just go forth ⁓ produce ⁓ whatever you need. So ⁓ that is essentially where we have got to. ⁓ ⁓ if you are really in the mood ⁓ you can actually see here there is a bit of a configurator. So a few months ago ⁓ I tried to ⁓ actually create marryme.com. You probably may have heard that from an old ⁓ episode and that was a song configurator built on anti-gravity. I got really far with it, then started to hit some issues. It didn't quite get live. ⁓ but this is going to be the latest incarnation of that. And essentially you will be able to come in and ⁓ just commission a song ⁓ or an album or indeed you know a string of albums. and eventually this will move from a request form to ⁓ basically just generate and so all of that is coming but I think just a really good showcase you know of where we have got to ⁓ this is the website just once again that Fable Five has built ⁓ and ⁓ all of this ⁓ ⁓ distributing music that this particular label San Carlos Records has already generated. ⁓ it is able to generate probably around ⁓ two albums a day. ⁓ the limitation is purely ⁓ the amount of time that it takes me to review ⁓ at the moment. I have no time at all, so I'm producing maybe two albums a week. ⁓ because that ten minutes it takes for me to review and upload them to SoundCloud truly is a point of friction. ⁓ but ⁓ let's see, I don't think we'll be able to overcome that entirely and Nor should we, ⁓ because that is really where the human in the loop comes in. So a multi-persona AI label broadcasting from the Blairy Coast, nine personas, one method, recordings you actually own, sounds, songs, albums, characters, and entire labels on request. So again, all of this built by AI, all of this run by AI. The AI has developed the strategy, and ⁓ really I'm just chaperoning it through the process. And so this is, I think, an excellent showcase of you know what the true capability of AI is today. I'm sure we will be able to go even further with it. But ⁓ any thoughts on that, Emil? Emile: It's looking great. You're packing a lot in there, Rolf. And as you mentioned, you kind of take the words right out of my mouth. It's a wonderful showcase of just showing what you can do, taking these frameworks to more sophisticated levels, getting your agents in, getting all the workflows in, getting the pipelines in. It's a perfect project, and I'm sure it's going to be welcomed by a lot of people. Nice job. Thanks for showing us. Ralph: And of course, you know, there's no expectation that any of this is actually going to work. All we're doing is we're really just testing and throwing everything that ⁓ AI can do at it and seeing where it goes. a kind of a fun project ⁓ for a Saturday. And I you people should go and have a play with it, see what's possible, certainly listen to the music. check out the radio station, it's a pretty cool little feature. And and actually that probably will be the next area that ⁓ when I get some time we will build out because we'll go back into video generation. And so the radio station ⁓ will actually also be looking at well what does an AI native radio station look like with AI native programming, AI native content, but really also trying to make sure that there are guardrails so that it's not producing slop. How do we get high quality stuff that actually is useful to us and that people value. so ⁓ no doubt I'll get some feedback via our various channels but ⁓ that's where we've got to this week looking at the monetization of this project I think I am still at $18 which is revenue earned over last year ⁓ and that is now off some 20,000 plays on SoundCloud. I guess I have one hit so far with one song that has garnered some four and a half thousand listens, many ⁓ thumbs ups and likes on on SoundCloud. but ⁓ you know ⁓ obviously not enough to make a living. although the cost of having all of this together is still reasonably low. It's kind of other than my time, you know, it's like in the hundreds. ⁓ of euros so not an expensive endeavor to get to this point just a lot of time investment well i say that obviously you need to have your various subscriptions which is probably the biggest cost so ⁓ Emile: Stick with it, it'll continue to grow and I love your very cool looking llama. I'd definitely party with him. He looks like he knows what he's doing. Ralph: Well, it's actually you know, because he's not a llama at all, he's an alpaca. So ⁓ yeah, yeah, the local farm ⁓ is called Al Pakistan. And right, are all very cool, at least ⁓ in terms of how they appear on my content. when you kind of hang out with them, they're a little bit more down to earth. but nonetheless, it's an exciting project to see, you know, how we can well, just kind of anthropomorph. Emile: I'll pack her, Ralph: anthropomorphies, these animals. ⁓ Emil, why don't we move on to something a little bit more serious? Emile: There you go. Yeah. No problem. And thanks for showcasing amazing work, Wolf. I wanted to circle back on GLM 5.2. been mentioning it in several last podcasts, because it is kind of gaining traction. And there does seem to be an unlikely rise with it. The Chinese open weight model has just achieved something that was supposedly impossible. It's beating Fable 5 at front end design, marking a big shift in in how we perceive global AI capability. To understand this, the AI Daily Brief compares us to DeepSeeks R1 women back in 2025. That was when a relatively unknown Chinese lab dropped a three reason model in Simon app. And suddenly the market had a complete meltdown in video. If you recall, lost a staggering 589 billion in market cap in a single day because investors thought the competitive mode had evaporated overnight. these Chinese open wake models follow predictable life cycle. They crushed the benchmarks as we know, generates insane hype for a week or two. ⁓ and then quietly disappear when real world usage proves them too fragile. The GLM 5.2 for now seems to be defying this trend. Predominant tech leaders are genuinely stunned by its coding capabilities, meaning it's actually surviving contact with real world developers. When we look at the specific data from designer reader for GLM 5.2, it's ranking first in design, which is pretty reliable news. specifically because it avoids notorious AI design flaws like those infamous early purple gradients if you would call and it's incredibly proficient at naturally calling complex dependencies like Chart.js and Tailwind.css. However, this capability does come with a serious cost constraint. The detailed outputs from GLM 5.2 results in 25 % more characters in lines of code and doubling the generation time. And running it locally at peak performance requires hardware that could cost you upwards north of $400,000. So it's not cheap. is a huge win, however, for a cloud AI layer. It means companies don't have to rent the most expensive state of the art model for every single task. They can now build diverse, cost-optimized architectures, leveraging powerful open-weight models or specific workflows while reserving the expense of frontier models for what when they truly needed. The era of simple two horse race in AR seems to be running its course. It's a good opportunity at least to start experimenting with these alternative architectures in a sandbox environment and see what works best for you. But I think, Rolf, there's just a common theme, like everything we've discovered, we've discussed with both GLM and Fable, you know, the same theme keeps playing through. People are starting to look at routes in more serious. that's just gonna become a developing trend it's just gonna mature ⁓ you all the clover that's happened with Fable 5 has kind of forced the hand ⁓ for path to be taken and it's now been carved out. I think that's gonna be the new reality of AI. Ralph: Yes, and ⁓ what's fascinating about it is that, you know, the US government can try to limit accessibility to the most highest performing models out there. China will do the same thing, maybe indeed they are doing the same thing. ⁓ but ⁓ of course these ⁓ open source models will continue to be available ⁓ and you know, it might start pushing people towards them. So that's kind of exciting. one thing that I will call out is ⁓ actually there is a really interesting stat. I'll probably get some of the specifics of it wrong, but you know, if you really look at what you are using, so when you are using something like a clawed code or a codex, the component that is the LLM is really small, less than two percent. So what's the other 98%? ⁓ Well, is known as the harness, and the harness is essentially ⁓ framework that has ⁓ it's like a tool belt really and reason why clawed code has been so successful is it had a very extensive tool belt. It could connect to many things, it could run many things, it had many skills, many plugins, and obviously this concept of MCP's ⁓ model context protocol which meant it could actually access many systems that previously were not accessible. And so ⁓ that incredible capability is really what has set clawed code apart. ⁓ the LLM is a small component of that. And when we talk about ⁓ GLM point two, we're really talking about you know the LLM component, not the harness around it. But you can also actually connect GLM five point two into clawed code. which means that ⁓ it can be the brain you can utilize the harness almost the same level as Claude Code. I ⁓ I that people talk about some limitations. and so that will become the new battleground, which kind of brings me to the cursor story because cursor ⁓ essentially ⁓ you is a harness ⁓ they're also building out their own LLM brain. And they have been purchased now by XAI for the princely sum of $60 billion. doubt a lot of that will be in various shares people now have ⁓ in the SpaceX company, but ⁓ probably a few new billionaires minted out of that deal. ⁓ and it's kind of interesting as well because I think it was a good move from Elon to actually buy Cursor. It means that XAI is partly back in the game. ⁓ but some of the critics have said ⁓ actually, you look at this, you know, Cursor was already declining. It went from something like 50% market share to about 25% market ⁓ share. it's you know. declining in a market that is rapidly expanding, which just goes to show ⁓ just how much competition has actually come to the fore. So you know, was Elon right to buy it? Was he right to spend sixty billion on it? ⁓ did he buy into something that is actually declining? was it a desperate move? I think all of those things are probably true. Are they criticisms? No, I think they're just the reality of the game. ⁓ what else was he going to do? ⁓ I think ⁓ XAI is sitting at something like two percent in terms of its overall user base. and so ⁓ you've got Gemini and OpenAI with Chat GPT leading ⁓ in terms of the ⁓ the lion's share of users but I think just from an internal of view that probably made sense for him to to have his have ⁓ have his own capability And so I'm not going to think too critically about that deal, but sixty billion did seem to be an awful lot of money. ⁓ although you know what, it just kind of ⁓ was created out of thin air when SpaceX actually did their IPO. So you know, it didn't cost him anything at the end of the day. He basically got it for free. Emile: Exactly. It's all relative ratios. Ralph: For the rest of us it's a big number. For Elon it's ⁓ it's just an uplift on ⁓ another good day's worth of SpaceX shares trading on stock. Emile: Roald, for what it's worth, just my personal take on that is think it's a perfect match in heaven or perfect marriage, so to speak, because they both had what the other needed. XAI had compute and curse. RL. They they they sitting on really valuable development data to apply the RL to their their new independent Kimmy train model, which is if not, it's already come, it's coming soon. and the proof will really be in the pudding when we see that what we've seen on the last training model it was already impressive. The benchmarks for what it's worth were saying that you know it's on par, closely matching ⁓ 4.8. ⁓ Which is a huge step change and improvement. I I know that the benchmarks aren't fully reliable, but regardless, the step change was there. So they're moving in the right direction. And a closing note too, you know, Groc really hasn't made a huge imprint ⁓ in the AI world. The penetration's been low subpar, I think it's somewhere between one to three percent. You called out out in a previous mod ⁓ episode, Rolf. So on balance, I I I think the needle's moving in the right direction. As you say, the IPO was so huge for SpaceX, it's a drop in the ocean for them. their bets. ⁓ gives them another card to play. ⁓ I I think it's quite an intelligent move personally by Elon, but ⁓ time will tell where the product goes. Ralph: Yes, and what's interesting is XAI has ⁓ now introduced a couple of new tools. So they've got their voice agent builder, a no code builder for Grok voice agents. So obviously 11 Labs was really the first mover in terms of AI voice technology, ⁓ and they have such a capable platform. ⁓ we've also seen that Mistral actually ⁓ about two months ago released the ability to essentially build your own 11 labs and now xai basically coming in with their voice agent builder so it's aimed at support sales ⁓ and for calls it's a point and configure voice agent that answers phones, qualifies leads, books appointments, ⁓ and ⁓ sing any songs. So I'm not quite sure if I'm gonna be able to use it for the San Carlos record label experiment but there you go. Mistral have also come out with ⁓ a optical recognition release. So it's ⁓ version four that they're on and this means basically able to extract documents at scale it's essentially a full enterprise ready AI solution for optical character recognition recognition so multiple format documents and extract structured data. and it's one of these things that they kind of see as a bit of an entry use case for any teams drowning in invoices, contracts and forms that they actually need to get their hands on. Now this is an interesting one. I I'm doing quite a bit of just work ⁓ on ⁓ looking extracting data from PDFs and Cloud Code does ⁓ such a good on it. I've not needed to you know go to Tools like Mistral for this OCR capability. enough, Quen this is Alibaba's LLM model, also has a very ⁓ capable ⁓ optical character recognition capability. So a little bit of kind of like also ran in terms of XAI and Mistral in that sense. are they doing enough to really pull users over to them? ⁓ they're kind of you know catching up or just playing. You know, in a space where solutions are already available. So not really sure what to make of that. and then the last new AI tool of the week is Liquid AI. ⁓ very interesting name. LFM 2.5 230 mil. So a 230 million parameter model pre-trained on 19 trillion tokens. Really, 19 trillion. now Liquid says that it beats models four times its size on data extraction. It runs anywhere. What does that mean? Well, it could run on your phone. So they're talking about these edge devices. ⁓ it could run on your calculator. And ⁓ and that's ⁓ interesting if that is true, because you know, over the years we've been talking about IoT. Does anybody talk about IoT? The internet of things? it just seems to have evaporated as a conversation. But when you st suddenly start having your fridge and your freezer with the access to the same level of intelligence as any frontier model, does it stop being a fridge or a freezer or a washing machine? ⁓ don't know what the answer to that is, but ⁓ I can imagine we are gonna get ⁓ kinds of weird and wonderful use cases. Emile: Well, you joke you joke, Rolf, but that does seem to be happening. There's several companies in on the valley that are starting to push edge computing to homes, taking advantage of the water supplies and the energy supplies and effectively not quite sure how they're doing it, but they're putting in a rack somehow into the household and sharing that back to the grid. So I think that theme is gonna continue to mature and play out. Ralph: Yeah, no doubt. I mean the robots are coming ⁓ you know at the end of the day, I would say if you have an EV, ⁓ electric vehicle, are driving a robot. it will soon driving you. And you know, ⁓ if we start to at the internet of things ⁓ and how we p these kind of edge device concepts, make every single electrical device in your home not just smart but you know super intelligent level. you know, who knows? But I still have a leak in my pool and the water is drained and my smart ⁓ know pool cleaner can't help me with that. So ⁓ life gets in the way of all of these things at some point. Emile: Ralph, ⁓ I just want to move on to a really cool story on the job front. ⁓ it's not for everyone, but ⁓ it's it's good for some of us. apparently will pay you eighty-five thousand dollars to learn AI. Seems almost too good to be true, and it's called Clawed Corpse, and it's just been launched. It's a 12-month paid fellowship for people at the very start of their careers, and it's 12-month job inside of a nonprofit. You'll be expected to build things, train staff, and leave behind something useful that actually works. There is an application process, a practical assessment, and interviews. You do need to be over 18 and have a US work rights and have less than two years of full-time work experience, but no degree is required though. So it's a pretty decent deal if you can get it, paying you to master the most in-demand skill on the planet right now. And handing you real world experience using it. It's a it's a legitimate, well-funded program aimed at giving early career people paid AI experience while helping nonprofits. So despite our youthful looks, I don't think it's for us, Rolf. But ⁓ you know, for those that it is applicable, check it out. It's it's a good offer. Ralph: Well, I think I can definitely tick the box. I'm ⁓ I'm over eighteen. ⁓ and I'm ready to join to sign me up. Emile: Yeah. Ralph: Although, you know, again, like we're all in the same place. This is what's so exciting about where this space is at. Everybody is new to this capability, whether you are just coming into the job market, whether you are a seasoned executive. and obviously, you know, maybe that is upset upsetting if you're in the latter part of your career, you feel like you shouldn't have to work this hard. ⁓ and of course, you know, ⁓ we've been working incredibly hard keeping up with all this news. learning new skills, learning new technologies at a rate that certainly ⁓ I have never encountered before. and it has been expansive you know, for me personally and ⁓ and for the people that I've been working with in this space. ⁓ just on a bit of an aside note, you know, I'm really seeing ⁓ the next site solutions that's you know I'm working with clients to really onboard this hardware that is now installed in stores and basically it is a wonderful label that I've actually got one just in front of me over here. It's a little label that goes on ⁓ ⁓ a tag and label can be read by various technologies in the store that basically can track every single item and where they move to and and this is starting to create some increase Incredible insight when you pair it with AI. So something like 4 billion data points per store per week. basically putting that all together, you know, was a really big data challenge. But in the age of AI, we're able to make incredible sense of the customer journeys that are evolving in the store ⁓ and actually driving some insight into human behavior, but also into how to make the the shopping experience better. ⁓ and that for me, you know, is just ⁓ a privilege and honor to to to work on. And so ⁓ I'm really quite bullish on some of what we've seen. But it is incredibly hard work. And and I think we can't take for granted just the sheer amount of effort that people are putting in. And also if I just jump on to kind of my next story, which is Probably a little bit of a an anecdotal story. ⁓ I've been reading a lot of articles and feedback on what means to be a developer in the of AI. And obviously over the last year, you know, we've had the rise ⁓ of vibe coding. ⁓ we have seen that tools like Cursor and ⁓ base you know, even Claude Code, ⁓ they and Pi, these tools have come to the fore ⁓ as kind of fundamental to the engineering experience of being a developer. ⁓ we have heard from Anthropic that ⁓ they have seen, you know, ⁓ probably about ⁓ six months ago, I think Boris Cherney, the the the founder of Claude Code, the inventor, if you like, was saying, you know, he has not written a line of code in about a month. So that was a ex post from some months back. ⁓ And all of this stuff is recent history. you know, the the world is literally in evolving in front of our very eyes. But what has it evolved to? ⁓ And what are saying is they have now become reviewers. ⁓ All they all day doing is ⁓ reviewing that AI has produced. Checking it for obvious errors, really the fun has gone out of the development process. And so while you've got the people with the experience that actually can review and make sense of what the AI has set out in the design document, ⁓ it's really celebrating their skills. And some of them don't like it. And some of them are experiencing a a new type of burnout, review burnout. basically. Where you are a passive participant the inventions ⁓ your AI coding platform. ⁓ ⁓ this has kind of meant that, you know, a number of developers are looking at this going, I'm not sure this is what I spent the last 10 years developing my skills for, just to become ⁓ glorified reviewer of output. ⁓ And obviously they're not feeling challenged. personally. I have a view on this. I think human creativity you know, is ⁓ abstracting up a level. it's much like what I see with the music. ⁓ you're not there to create a hook or a sound, ⁓ or even a lyric. You're there to define how you want people to feel, the concept that you want the world to consume. And so that becomes Emile: Yeah. Ralph: Essentially your instrument that AI then interprets into a person persona, a song, an album, a series of albums. And so I think it changes our perception what the instrument is ⁓ where human creativity needs to play out. However, this is my final point on this, I just can't get away from the fact that, you know, engineers, developers that have spent years and years refining their skills of finding that those skills are not relevant and that they're really not enjoying the part that remains. while they are really the only people that can evaluate ⁓ what AI is outputting, they are no longer the ideal people to actually be doing that work. And so I think we've got a bit of a conundrum over here where people myself that ⁓ are system operating model designers, you know, able to put high level concepts together, are really enjoying the creative release of AI interpreting our ideas. developers are just finding it that it's mind numbing review work and they don't want any of it. ⁓ And leading to a new kind of burnout. Emile: Yeah, I t I totally agree. I'm seeing it myself, Rolf. ⁓ I think there's even a term for it. We were discussing it, ⁓ instead of bot sitting, it's called bot shitting. ⁓ it's effectively exactly what you're calling out. and in addition to that, that exacerbated feeling ⁓ of having their worth or just policing over something is also creating a reduction in agencies, agency with these people, which is not ideal. ⁓ so it's it's effectively rising the attrition rate in jobs and people are looking for alternative placements just because they're bored or they don't don't feel like they they're offering the value or as you called out rightly, you know, it's something they've trained years for. So ⁓ I think we do need to watch this one I'm I'm not quite sure how it plays out. And all the original doommongering from the start of the year with the job apocalypse, that also seems to be a game of two halves. Is ⁓ if anything, the human in the loop is coming back and trending a lot harder now. And there have actually been an increase in jobs, but it depends how you cut it up. If you divorce and you dissect and you take finance development work specifically, those jobs are reducing. But holistically, if you look at the whole job market, that seems to be increasing. So it's penetrating in a certain direction. Maybe that displacement will start widening up. I'm not sure. But ⁓ yeah, there's just there's two parts of the story. It's not as black and white as you might see in the papers. Ralph: Definitely not. I think there's a lot ⁓ you know that is incredibly positive. I was also reviewing a nice ⁓ news article this week about Ford that actually had put in ⁓ ⁓ of AI solutions to review the quality ⁓ the the line ⁓ and have started hiring grey haired engineers. Back into the production line to work in the factories. So guess what? They're going to have to take the canteen back. It was used as a battery storage bank for the robots. Now it's going to be used to serve humans tea once again. I've completely made that up. the point is that you know we're bringing humans back into the loop here ⁓ in a significant way. ⁓ and I think all of this still has quite a long way to play out. But Emile: Hey. Ralph: You know, the world of work is shifting and changing in really kind of quite meaningful ways. And so we should all just take a moment to appreciate just how significant this moment is in history as we're all kind of trying to find ⁓ a new way. But I have basically finished most of my stories and I've kind of been furiously ⁓ scrolling through my research over here because I As you know, we've also been testing ⁓ over the weeks whether AI can actually generate a joke, but I'm actually not seeing my joke generator. It didn't Emile: That's fine as as you look through Ralph, I've got ⁓ just a couple of very small buy pieces of news I thought was interesting. from Theo, he's saying he you know, he's well quite well connected in the market. He's saying it's insane how many people you know mostly use clawed models because their company is on AWS and they're using bedrock, so it's by default. I thought that was quite interesting. And another news item, the US Secretary of War, Pete Hegseth just appointed Mark Adreessen to the Defense Policy Board. if you don't know him, he's part of a 16Z podcast group now officially ⁓ Pentacon on national security. And on topic of Mark Andreessen, he's also declaring on X that Dr. G Chat GBT is already a better doctor than 99% ⁓ doctors. However, comments on X, remain cautiously skeptical on it. I'm not sure I'd want to be putting that doctor on ⁓ you know, in in in ports and surgeries just yet. But yeah. There you go. Ralph: Yeah, so that's our route. you know, from podcasts to ⁓ security advisors. it's nice to see Mark blazing a trail for us. But ⁓ I did find a couple of jokes and big health warning on these. They are horrendous but somewhat amusing. So ⁓ let's have a look. So ⁓ first joke actually what just to start, it says There was a rare disease story intentionally kept out of the jokes. which is which I really want to know what that joke was about the rare disease story. But I've asked it obviously to keep it civil ⁓ and not be mean to anybody. but it came out with Google paid two point seven billion dollars to bring Noam Shazir back. So this was a guy that left Google. I think he I don't I'm not sure where he went. First, but I think it was OpenAI. and years later he went back to OpenAI. So somewhere in Google's finance department, that's filed under temporary staffing. ⁓ yeah, really interesting ⁓ that ⁓ paid so much money to get this guy back, and then he's headed off again. I wonder what that's cost him. But I mean, incredible that you know these people are being hired and Emile: Yeah. Ralph: Being turned into billionaires themselves multiple times by multiple companies. It just shows you how hot the market is for the top talent. next joke: SpaceX just pulled off the biggest startup acquisition in history. They bought Cursor, the AI code editor, for $60 billion. A rocket company paid $60 billion for autocomplete. What? It's not ironic and it's ⁓ left field enough for me. So the third one, the best open source coding model on earth, shipped this week, free MIT licensed, runs on your laptop, and it's Chinese. Somewhere a pricing team at a US lab is explaining to their board why the competition doesn't send invoices. That actually ⁓ it's the best laugh we've had. Emile: That's pretty good. Ralph: and then ⁓ just the final one over here. SpaceX owns Cursor. It also rents its data center to Anthropic, to Google, and this week 6.3 billion to Reflection AI. every major now runs on Elon's computer. They're not competitors, they're tenants. All right, that was more of a fact than than a funny but ⁓ there you go. So don't worry, people, we are safe. We have not yet achieved true super intelligence. ⁓ cause you know, I don't think any of those can qualify for me to go on stage at the comedy club and get an entire audience rolling on the floor with laughs. Anything anything to take us home, Emile? Emile: No, that's it from me, but ⁓ I appreciate your commended effort for keeping up the good fight with the jokes. One day we'll get there. Ralph: Yes, yes. And of course, know, ⁓ week I think we will probably do a bit of a model shootout, just exactly where we're at. ⁓ there will be more on the San Carlos Records experiment. can we actually get two twenty dollars in revenue? And I think will be an interesting project that ⁓ folks will be able to try out. which would have been released this week, but the AI ⁓ able to produce the guide. And that is, I have built a version of Andre Kopathy's self-recursive ⁓ knowledge And that's really interesting tool that I think is a fundamental staple for anybody. ⁓ you know, if you are in the world of work or you have a hobby, then what you want to do is keep up with the latest and the greatest news. And this self-recursive knowledge base does that. ⁓ I have a secret to share because I have actually extended it ⁓ go from an information model into an applied knowledge model. So ⁓ you will be able to give me some feedback, Emil, on how well it has done that. It is really pushing the envelope, and some might say it's even the bleeding edge ⁓ of. management, but ⁓ what else would you expect from a show called the Bleeding Edge Podcast? Emile: Look forward to it. Ralph: Well, thank you very much. It's been a great ⁓ week. Thanks for your contribution, Emil, and we'll see you on the edge. Emile: Adios.